Time Series Analysis of Aviation Data

Transcription

1 Time Series Analysis of Aviation Data Dr. Richard Xie February, 2012

2 What is a Time Series A time series is a sequence of observations in chorological order, such as Daily closing price of stock MSFT in the past ten years Weekly unemployment claims in the past 2 years Monthly airline revenue passenger miles in the past ten years Time series analysis is useful when No other data available System too complicated to model in detail

3 Where to Get the Data?

4 What Information Are You Interested In? How the data changes from month to month, year to year? Any trend? How fluctuated the curve is? Any seasonal effects? Any unusual years/months which have significantly small or large number? Can we forecast future value based on the time series?

5 Let s Work on the Data But first, what tool will you use? Pencil and quadrille pad (or back of an envelope) Excel Matlab, Mathematica, Maple SAS, SPSS, STATA, R ROOT, PAW, KNIME, Data Applied, etc. Others

6 Use R! R is free R is a language, not just a statistical tool R makes graphics and visualization of the best quality A flexible statistical analysis toolkit Access to powerful, cutting-edge analytics A robust, vibrant community Unlimited possibilities

7 Where to Download R To download R Go to Choose a CRAN Mirror, such as Click the link to download R according to your operating system (Linux, MacOS X, or Windows) Or Enhanced Version of R Distributed by 3 rd Party Revolution R Community ( Free academic version of Revolution R Enterprise ( CRAN: Comprehensive R Archive Network

8 R References An Introduction to R, W.N.Venables, D.M.Smith and R Development Core Team More Documents/Tutorials, go to

9 Start an R Project Recommend using RStudio as the console ( Create a project folder for storing R scripts, data, etc. e.g. C:/Users/xie/Documents/SYST460/R projects/airline_timeseries Open RStudio, navigate to the project folder 1. Use getwd() to find out the current working directory 2. Click and select the project folder 4. Current working directory is changed 3. Click here

10 Work on Data, Finally! Use Ctr+Shft+N to create a new script > rm(list=ls(all=true)) to clear the existing variables in workspace, if any > rpm = read.csv("system Passenger - Revenue Passenger Miles (Jan Oct 2011).csv") Use Ctr+Enter to run the current line or selection

11 What the Data Looks Like? > ls(rpm) > plot(rpm), equivalent to > plot(rpm$total~rpm$yyyymm)

12 Plot It As A Time Series > rpm.ts = ts(as.numeric(rpm$total), start = c(1996,01),freq=12) > plot(rpm.ts,ylab='revenue Passenger Miles') Commands to check properties of rmp.ts

13 Trends and Seasonal Variation > layout(1:2) > plot(aggregate(rpm.ts)) > boxplot(rpm.ts~cycle(rpm.ts)) Overall trend of increasing over years Max value Upper Quartile Median Lower Quartile Min value

14 Window Function Extract a part of the time series between specified start and end points > rpm.feb <- window(rpm.ts, start = c(1996,02), freq = TRUE) > rpm.aug <- window(rpm.ts, start = c(1996,08), freq = TRUE) > mean(rpm.feb)/mean(rpm.aug)

15 Modeling Time Series - Notations

16 Modeling Time Series Decomposition Models

17 Time Series Decomposition in R > rpm.decom=decompose(rpm.ts) > plot(rpm.decom)

18 Auto-Correlation

19 AutoCorrelation and Correlogram > rpm.acf=acf(as.numeric(rpm.ts),lag.max=40)

20 Correlogram after Decomposition >acf(as.numeric(rpm.decom$random),na.acti on=na.omit,lag.max=40)

21 Regression Trends: stochastic trends, deterministic trends Deterministic trends and seasonal variation can be modeled using regression Deterministic trends are often used for prediction Time series regression differs from standard regression as time series tends to be serially correlated

22 Linear Models

23 Fit A Linear Regression Model > fit = lm(rpm.ts ~ time(rpm.ts)) > plot(rpm.ts, type="o", ylab="rpm") > abline(fit)

24 Diagnostic Plots > hist(resid(fit)) > acf(resid(fit)) > AIC(fit)

25 Linear Model with Seasonal Variables

26 Linear Model with Seasonal Variables - R > Seas = cycle(rpm.ts) Gives the positions in the cycle of each obsv. > Time = time(rpm.ts) Creates the vector of times at which rpm.ts was sampled > rpm.lm = lm(rpm.ts~0+time+factor(seas)) Fit rpm.ts to the linear model with seasonal variables

27 Take a Look at rpm.lm > summary(rpm.lm) Model Residuals from fitting Coefficients

28 Correlogram of rpm.lm Residual > acf(resid(rpm.lm),lag.max=40) It indicates strong positive-autocorrelation Residuals are not pure random numbers, so it should be further modeled

29 How Random is Random? - White Noise A time series {w t : t = 1, 2,..., n} is discrete white noise (DWN) if the variables w 1,w 2,...,w n are independent and identically distributed with a mean of zero. This implies that the variables all have the same variance σ 2 and Cor(w i,w j ) = 0 for all i j. If, in addition, the variables also follow a normal distribution (i.e., w t N(0, σ 2 )) the series is called Gaussian white noise.

30 Simulate White Noise in R > set.seed(1) > w = rnorm(100) > plot(w, type = "l") > acf(w)

31 Random Walk Let {x t } be a time series. Then {x t } is a random walk if x t = x t 1 + w t where {w t } is a white noise series. Substituting x t 1 = x t 2 +w t 1 and then substituting for x t 2, followed by x t 3 and so on gives: x t = w t + w t 1 + w t In practice, the series will start at some time t = 1. Hence, x t = w 1 + w w t

32 Simulate A Random Walk in R > x <- w <- rnorm(1000) > for (t in 2:1000) x[t] <- x[t - 1] + w[t] > plot(x, type = "l") > acf(x) Your plot will be different for sure!

33 Stationarity

34 Auto-Regressive (AR) Models The series {x t } is an autoregressive process of order p, abbreviated to AR(p), if x t = α 1 x t 1 + α 2 x t α p x t p + w t where {w t } is white noise and the α i are the model parameters with α p 0 for an order p process. The model is a regression of x t on past terms from the same series; hence the use of the term autoregressive.

35 Fit an AR Model > res.ar=ar(resid(rpm.lm),method="ols") > res.ar x t = x t x t x t 3 + w t > acf(res.ar$res[-(1:3)]) OLS: Ordinary Least Square

36 Moving Average Model

37 Simulate an MA Process > set.seed(1) > b <- c(0.8, 0.6, 0.4) > x <- w <- rnorm(1000) > for (t in 4:1000) { for (j in 1:3) x[t] <- x[t] + b[j] * w[t - j] } > plot(x, type = "l") > acf(x)

38 Plot Results First 3 lags with significant correlation Simulated MA Process Correlogram of the Simulated MA Process

39 ARMA Model

40 Fit An ARMA Model > rpm.arma=arima(rpm.ts,order=c(1,0,1)) > acf(as.numeric(rpm.arma$resid)) Strong seasonal information left in residuals

41 ARIMA and SARIMA Model A time series {x t } follows an ARIMA(p, d, q) process if the d th differences of the {x t } series are an ARMA(p, q) process SARIMA is Seasonal ARIMA model which extends ARIMA model with seasonal terms

42 Fit A SARIMA Model > rpm.arima=arima(rpm.ts,order=c(1,1,1), seas=list(order=c(1,0,0),12)) First c(1,1,1): AR(1), first-order difference, MA(1) Second c(1,0,0): seasonal terms on AR process, frequency 12 > acf(as.numeric(rpm.arima$resid))

43 Forecast Predicting future values of a time series, x n+m, using the set of present and past values of the time series, x={x n, x n-1,, x 1 } The minimum mean square error predictor of n x n+m is x E( x x) n m predict(model, newdata) method model: a model object used for prediction newdata: value of explanatory variables n m

44 Forecast in R > new.t <- seq( , len = 2 * 12, by = 1/12) > new.dat <- data.frame(time = new.t, Seas = rep(1:12, 2)) > rpm.pred=ts(predict(rpm.lm, new.dat)[1:24],start=c(2011,11),freq=12) > ts.plot(rpm.ts,rpm.pred,lty=1:2)

45 Plot of Forecast

46 Forecast with SARIMA Model > ts.plot( cbind( window(rpm.ts,start = c(1996,1)), predict(rpm.arima,48)$pred ), lty = 1:2)

47 Homework Download and Install R with Rstudio Read An Introduction to R Download System Passenger - Revenue Aircraft Miles Flown (000) (Jan Oct 2011) data from BTS Read the data into R using Rstudio Create a time series plot of the data, and plot its auto-correlation correlogram Decompose the time series and save the plot

48 Homework (Ctd.) Construct a linear regression model without seasonal factors, and plot the correlogram of the model s residual data Construct a linear regression model with seasonal factors, and identifies the characteristics of the model residual. Fit an AR model to the model residual of the above model Forecast the time series data into next 24 months using the seasonal model

49 Exam Questions What are the data elements in a time series? What does auto-correlation mean? What are white noise and random walk? What are stationary models and non-stationary models?